A native of the British language asks if he has spelled a British name right...
I could give a basic exposition of the theory of Bayesian networks. But I suspect you have an application-driven question in mind, and there a different angle might be better.
At present, I am not Bayesianing anything. I am considering, however, making myself a kind of agent program, which would scan through mailing lists, rss feeds, google news, and other text sources, and keep items for me to look at, learning over time what I'm interested in and not interested in. I figure that if spam filters can learn what spam is from examples, an agent can learn what I like from examples.
I was afraid of that. The learning over time aspect is something I'm not too well versed of -- as yet, at least. But it is being done, and with well-known methods. So there should be no big issues there.
So, what you are going to have is two modules: 1) the Bayesian network doing classification of items, and 2) the adaptation of the network based on user decisions. Part 2 you need to find from other sources, and naturally what I'm lecturing here about part 1 is given without warranty, express or implied, of suitability to any purpose or even consonance with generally perceived reality.
In a Bayesian decision network, you implement the Bayes' posterior probability formula in a structured manner. First you need to decide what your output categories are. At simplest for your application, this could be two items: "bookmark" and "disregard".
Then you need to discover the input features. This is more an art or craft than engineering or science. When you know what the features are that you are basing your decision on, and what value combinations lead to what decision, you construct your network.
Mathematics seems to require that the network is preferably a tree, or at least an acyclic graph. The nodes of the graph are the various features. The arcs of the graph describe the dependencies between the variables. Now we have the structure of the graph.
Then we need to assign the a priori probabilities. This is where the adaptation algorithms work. The feature set is fixed, but the value sets allowed for each feature can be adapted, as well as the conditional probabilities between the dependency items.
And then you compute. From an item under examination, you extract the features, you propagate their weights through the network, resulting in a conditional a posteriori probability that the item is or is not interesting.
Don't know anything about Bayesian, but... I just *had* to tell you that it's been snowing all night and continuing this morning. Yes, we even have snow on Mar. 17th. >,-) And it's getting so, whenever I see snow, I think of you.
![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
![[identity profile]](https://www.dreamwidth.org/img/silk/identity/openid.png){kind=small}
From:
no subject
I could give a basic exposition of the theory of Bayesian networks. But I suspect you have an application-driven question in mind, and there a different angle might be better.
So. What are you Bayesianing?
From:
no subject
From:
no subject
So, what you are going to have is two modules: 1) the Bayesian network doing classification of items, and 2) the adaptation of the network based on user decisions. Part 2 you need to find from other sources, and naturally what I'm lecturing here about part 1 is given without warranty, express or implied, of suitability to any purpose or even consonance with generally perceived reality.
In a Bayesian decision network, you implement the Bayes' posterior probability formula in a structured manner. First you need to decide what your output categories are. At simplest for your application, this could be two items: "bookmark" and "disregard".
Then you need to discover the input features. This is more an art or craft than engineering or science. When you know what the features are that you are basing your decision on, and what value combinations lead to what decision, you construct your network.
Mathematics seems to require that the network is preferably a tree, or at least an acyclic graph. The nodes of the graph are the various features. The arcs of the graph describe the dependencies between the variables. Now we have the structure of the graph.
Then we need to assign the a priori probabilities. This is where the adaptation algorithms work. The feature set is fixed, but the value sets allowed for each feature can be adapted, as well as the conditional probabilities between the dependency items.
And then you compute. From an item under examination, you extract the features, you propagate their weights through the network, resulting in a conditional a posteriori probability that the item is or is not interesting.
From:
Snow...